Molecule generation method and apparatus, electronic device, and storage medium

Abstract

The application provides a molecule generation method and device, electronic equipment and a storage medium, wherein the method comprises: obtaining a first transcriptome state feature before cell drug disturbance and a second transcriptome state feature after cell drug disturbance; determining a molecular condition vector based on the first and second transcriptome state features; inputting the molecular condition vector into a conditional molecule generation model to obtain a molecular structure; by converting the functional difference at the phenotype level into a constraint condition for chemical molecule design, the conditional molecule generation model can be guided to generate a molecular structure capable of causing a specific cell disturbance response, and precise de novo molecule design can be realized in a target unknown or multi-pathway complex disease scenario. The method takes the cell phenotype change as the generation condition, effectively bridges the modality gap between biological signals and chemical structures through a double-view alignment strategy, realizes reliable mapping from functional characterization to structural design, and significantly improves the accuracy of de novo drug molecule design based on transcriptome guidance.

Description

Technical Field

[0001] This invention relates to the field of deep learning-driven drug molecule design technology, and in particular to a molecule generation method, apparatus, electronic device, and storage medium. Background Technology

[0002] Currently, structure-based drug design (SBDD) dominates the field of drug discovery. SBDD relies on the "lock and key" principle, using the three-dimensional structure of the target protein to guide the generation of high-affinity ligands. On the other hand, cellular perturbation transcriptomics provides a comprehensive means of capturing snapshots of cellular functional states. Large-scale perturbation databases provide massive gene expression profiles under chemical or genetic perturbations. Existing transcriptomics-based machine learning methods are mainly used to solve positive prediction problems, i.e., integrating chemical structure and baseline state to predict the cellular response of known compounds at the single-cell or tissue level. However, when processing large-scale and sparse single-cell transcriptome data, existing methods in computational biology for generating and predicting baselines typically employ pseudo-bulk analysis techniques, i.e., aggregating heterogeneous single-cell populations into a single, unified representation vector through simple averaging. In terms of molecular generative architectures, deep molecular design has evolved from SMILES (Simplified Molecular Input Line Entry System) sequence models to graph-based methods that preserve molecular topology.

[0003] However, the aforementioned existing technologies still face several key bottlenecks and limitations in the development of functional drugs for complex diseases. First, the SBDD method, limited by its reductionist paradigm, is difficult to apply in situations where the target structure is unknown or the disease involves multiple pathway network disorders rather than a single target, thus restricting its application in systemic diseases. Second, transcriptome-based methods suffer from significant information asymmetry; most studies only use transcriptome features as prediction targets rather than as conditional signals for the reverse generation of new molecules, resulting in the underutilization of the potential of large-scale perturbation data. Even the few attempts to guide molecular generation based on transcriptomes still rely on explicit statistical features, lacking robust cross-modal alignment and interaction mechanisms, which easily leads to information loss. Third, cross-modal mapping from biological transcriptomes to chemical molecular structures presents fundamental challenges. The two differ significantly in information density, structural representation, and domain priors, making it difficult for direct conditional generation to maintain stability and chemical rationality. Summary of the Invention

[0004] This invention provides a molecular generation method, apparatus, electronic device, and storage medium to address the shortcomings of existing technologies in drug development for complex diseases, such as strong target structure dependence, insufficient utilization of transcriptome information, and insufficient stability of cross-modal generation.

[0005] This invention provides a method for generating molecules, comprising the following steps.

[0006] Obtain the first transcriptome state characteristics of the cells before drug perturbation and the second transcriptome state characteristics of the cells after drug perturbation; The first transcriptome state features and the second transcriptome state features are input into the dual-view molecular domain model to obtain the molecular condition vector output by the dual-view molecular domain model. The molecular condition vector is input into the conditional molecular generation model to obtain the molecular structure corresponding to the cell perturbation response output by the conditional molecular generation model. The dual-view molecular domain model includes a molecular graph autoencoder, an activity representation extraction module, and a fusion module. The molecular graph autoencoder is used to determine perturbation transcriptome features characterizing cellular perturbation responses based on the first transcriptome state features and the second transcriptome state features. The activity representation extraction module is used to determine global topological representation and local activity representation based on the perturbation transcriptome features. The global topological representation is used to characterize the global topological legitimacy of the molecule, and the local activity representation is used to characterize the local functional activity of the molecule. The fusion module is used to fuse the global topological representation and the local activity representation to obtain the molecular conditional vector.

[0007] According to a molecular generation method provided by the present invention, the step of determining perturbation transcriptome features characterizing cellular perturbation responses based on first transcriptome state features and second transcriptome state features includes: The first transcriptome state features and the second transcriptome state features are subjected to feature interaction to obtain the first interaction feature and the second interaction feature; The perturbation transcriptome features are obtained by fusing the first interaction feature and the second interaction feature.

[0008] According to a molecular generation method provided by the present invention, the dual-view molecular domain model is obtained by iteratively executing the following steps until a preset iteration termination condition is met: A sample perturbation transcriptome feature set and an initial dual-view molecular domain model are obtained; the initial dual-view molecular domain model includes an initial molecular graph autoencoder and an initial activity representation extraction module; the sample perturbation transcriptome feature set includes multiple sample perturbation transcriptome features, the sample perturbation transcriptome features include a first sample perturbation transcriptome feature and a second sample perturbation transcriptome feature; the initial molecular graph autoencoder includes an encoder and a decoder; the initial activity representation extraction module includes an initial local activity representation extraction module; The sample perturbed transcriptome features are input into the initial molecular map autoencoder, the encoder encodes the sample perturbed transcriptome features to obtain the latent code, and the decoder decodes the latent code to obtain the molecular map; Based on the molecular graph and the latent encoding, the global topological alignment loss is determined; Obtain the target Morgan fingerprint corresponding to the perturbed transcriptome features of the sample; The sample perturbation transcriptome features are input into the initial local activity representation extraction module to obtain the predicted continuous fingerprint representation output by the initial local activity representation extraction module; Based on the predicted continuous fingerprint representation and the target Morgan fingerprint, determine the local bioactivity alignment loss; The target loss is determined based on the global topology alignment loss and the local bioactivity alignment loss; Based on the target loss, update the model parameters of the initial dual-view molecular domain model.

[0009] According to a molecular generation method provided by the present invention, determining the local bioactivity alignment loss based on the predicted sequential fingerprint representation and the target Morgan fingerprint includes: Obtain a positive sample mask and a negative sample mask; wherein, the positive sample mask indicates the valid positions in the target Morgan fingerprint, and the negative sample mask indicates the missing positions in the target Morgan fingerprint; Based on the positive sample mask, the predicted continuous fingerprint representation, and the target Morgan fingerprint, a first loss is determined; The second loss is determined based on the negative sample mask and the predicted continuous fingerprint representation; Based on the first loss and the second loss, determine the sparse regression loss; Based on the predicted continuous fingerprint representation and the target Morgan fingerprint, determine the contrast loss; The local bioactivity alignment loss is determined based on the sparse regression loss and the contrast loss.

[0010] According to a molecular generation method provided by the present invention, determining the global topological alignment loss based on the molecular graph and the latent encoding includes: Based on the molecular diagram and the latent encoding, the lower bound loss of evidence is determined; Based on the mean of the sample perturbation transcriptome feature set and the mean of the latent code corresponding to each sample perturbation transcriptome feature, as well as the variance of the sample perturbation transcriptome feature set and the variance of the latent code corresponding to each sample perturbation transcriptome feature, the feature distribution alignment loss is determined. The global topology alignment loss is determined based on the evidence lower bound loss and the feature distribution alignment loss.

[0011] According to a molecular generation method provided by the present invention, obtaining a first transcriptome state characteristic of cells before drug perturbation and a second transcriptome state characteristic of cells after drug perturbation includes: Acquire first and second single-cell transcriptome data; The first single-cell transcriptome data and the second single-cell transcriptome data are projected onto a low-dimensional manifold space to obtain the first projected cell embedding and the second projected cell embedding. Based on the cell cycle of the cells, the first projected cell embedding and the second projected cell embedding are grouped to obtain the first structured feature matrix and the second structured feature matrix; Local pooling aggregation is performed within each group of the first structured feature matrix and the second structured feature matrix, and the results of local pooling aggregation are hierarchically sampled and recombined to obtain the first transcriptome state features and the second transcriptome state features.

[0012] According to a molecular generation method provided by the present invention, the step of determining perturbation transcriptome features characterizing cellular perturbation responses based on the first transcriptome state features and the second transcriptome state features further includes: Determine the similarity coefficient between the perturbed transcriptome features and the fingerprints of each compound in a pre-defined compound fingerprint database; Candidate compounds are selected from the compound fingerprint database based on the similarity coefficient.

[0013] The present invention also provides a molecule generation apparatus, comprising the following modules: The acquisition module is used to acquire the first transcriptome state characteristics of the cells before drug perturbation and the second transcriptome state characteristics of the cells after drug perturbation. The input module is used to input the first transcriptome state features and the second transcriptome state features into the dual-view molecular domain model to obtain the molecular condition vector output by the dual-view molecular domain model. The generation module is used to input the molecular condition vector into the conditional molecule generation model to obtain the molecular structure corresponding to the cell perturbation response output by the conditional molecule generation model. The dual-view molecular domain model includes a molecular graph autoencoder, an activity representation extraction module, and a fusion module. The molecular graph autoencoder is used to determine perturbation transcriptome features characterizing cellular perturbation responses based on the first transcriptome state features and the second transcriptome state features. The activity representation extraction module is used to determine global topological representation and local activity representation based on the perturbation transcriptome features. The global topological representation is used to characterize the global topological legitimacy of the molecule, and the local activity representation is used to characterize the local functional activity of the molecule. The fusion module is used to fuse the global topological representation and the local activity representation to obtain the molecular conditional vector.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the molecular generation method as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the molecular generation method as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the molecular generation method as described above.

[0017] The molecular generation method, apparatus, electronic device, and storage medium provided by this invention acquire the first transcriptome state features of cells before drug perturbation and the second transcriptome state features of cells after drug perturbation; input the first and second transcriptome state features into a dual-view molecular domain model to obtain the molecular condition vector output by the dual-view molecular domain model; input the molecular condition vector into a conditional molecular generation model to obtain the molecular structure corresponding to the cell perturbation response output by the conditional molecular generation model. By transforming phenotypic functional differences into constraints for chemical molecular design, the conditional molecular generation model can be guided to generate molecular structures that can induce specific cell perturbation responses. This solves the technical problem of traditional drug design relying excessively on the three-dimensional structural information of the target, and enables accurate de novo molecular design even in scenarios with unknown targets or complex multi-pathway diseases. Meanwhile, this method uses molecular condition vectors characterizing changes in cell phenotypes as generation conditions, overcoming the information asymmetry problem in existing transcriptomics methods that only use phenotypes as prediction targets. It also effectively bridges the modal gap between biological signals and chemical structures through a dual-view alignment strategy, solving the problems of information heterogeneity and semantic gap between biological transcriptomes and chemical structures. This enables a reliable mapping from functional characterization to structural design, significantly improving the feasibility, accuracy, and generalization ability of transcriptome-guided de novo drug molecule design. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is one of the flowcharts of the molecular generation method provided by the present invention.

[0020] Figure 2 This is the second schematic diagram of the molecular generation method provided by the present invention.

[0021] Figure 3 This is a schematic diagram of the molecular generation device provided by the present invention.

[0022] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] The terms "first," "second," etc., used in this invention are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and that the objects distinguished by "first," "second," etc., are generally of the same class.

[0025] This invention provides a molecular generation method. The core idea of ​​this method is to formalize the drug discovery process as a generative inverse problem. Instead of predicting cellular phenotypic changes based on known molecular structures, it reverses this process by generating novel molecular structures that can induce desired cellular phenotypic changes (manifested as alterations in transcriptome state). This method eliminates dependence on the three-dimensional structure of specific targets, allowing for drug design directly from system-level functional outcomes. It is particularly suitable for complex diseases with unknown targets or those caused by dysregulation of multiple pathway networks. Figure 1 This is one of the schematic flowcharts of the molecular generation method provided by the present invention, such as... Figure 1 As shown, the method includes the following steps: Step 110: Obtain the first transcriptome state characteristics of the cells before drug perturbation and the second transcriptome state characteristics of the cells after drug perturbation.

[0026] Specifically, firstly, the first transcriptome state characteristics of the cells before drug perturbation and the second transcriptome state characteristics of the cells after drug perturbation can be obtained.

[0027] There are several ways to obtain the first and second transcriptome state characteristics. One possible approach is through direct determination via biological experiments. For example, a cell line can be cultured in vitro and divided into experimental and control groups. Before applying a drug (i.e., the perturbation source), the cells are sampled, and their gene expression profiles are measured using high-throughput sequencing technologies, such as single-cell RNA sequencing (scRNA-seq) or bulk RNA sequencing (bulkRNA-seq). This gene expression profile data, after processing, can serve as the first transcriptome state characteristic. Subsequently, the experimental group cells are perturbed by the drug, and after a certain period, the cells are sampled and sequenced again. The resulting gene expression profile data, after processing, serves as the second transcriptome state characteristic. Another possible approach is to obtain the data from public or private databases. For example, existing paired gene expression data for specific cells before and after perturbation with a specific drug can be retrieved and downloaded from large perturbation databases.

[0028] Here, the first and second transcriptome state features are used as data representations reflecting the overall gene activity state of the cell before and after perturbation. In some embodiments, the first and second transcriptome state features can be high-dimensional numerical vectors representing the expression levels of the entire genome or a specific subset of genes. Each dimension of the high-dimensional numerical vector corresponds to a gene, and its value can be the original sequencing read, normalized expression value, or other normalized value. In other embodiments, to address the common problems of high dimensionality, high sparsity, and technical noise in single-cell sequencing data, after obtaining the original expression profile, a preprocessing module can be used to project the high-dimensional, sparse original expression profile into a lower-dimensional latent manifold space with denser biological information, resulting in a denoised and aggregated low-dimensional vector as the transcriptome state feature. Thus, the first and second transcriptome state features can more robustly and accurately reflect the functional state of the cell.

[0029] Step 120: Input the first transcriptome state features and the second transcriptome state features into a dual-view molecular domain model to obtain the molecular condition vector output by the dual-view molecular domain model; the dual-view molecular domain model includes a molecular graph autoencoder, an activity representation extraction module, and a fusion module; the molecular graph autoencoder is used to determine perturbation transcriptome features characterizing cellular perturbation responses based on the first transcriptome state features and the second transcriptome state features; the activity representation extraction module is used to determine global topological representation and local activity representation based on the perturbation transcriptome features; the global topological representation is used to characterize the global topological legitimacy of the molecule, and the local activity representation is used to characterize the local functional activity of the molecule; the fusion module is used to fuse the global topological representation and the local activity representation to obtain the molecular condition vector.

[0030] Specifically, after obtaining the first transcriptome state features and the second transcriptome state features, the first transcriptome state features and the second transcriptome state features can be input into the dual-view molecular domain model to obtain the molecular condition vector output by the dual-view molecular domain model.

[0031] Here, the dual-view molecular domain model is a trained deep learning model whose core function is to translate biological signals representing cellular responses into a set of chemical property constraints to guide molecular generation. The "dual view" refers to the model's ability to resolve and encode the desired molecule from two complementary chemical perspectives: global topological legitimacy and local functional activity, thereby ensuring that the final generated molecule is both structurally sound and functionally specific.

[0032] The molecular condition vector is a high-dimensional numerical vector used to reflect and encode the global and local chemical properties that a molecule should possess to cause a transition from a first transcriptome state characteristic to a second transcriptome state characteristic. The molecular condition vector will serve as a guiding signal for subsequent conditional molecule generation models.

[0033] Here, the dual-view molecular domain model includes a molecular graph autoencoder, an activity representation extraction module, and a fusion module. The molecular graph autoencoder is used to determine perturbed transcriptome features characterizing the cellular perturbation response based on first and second transcriptome state features. This feature extraction process aims to accurately separate and capture the net transcriptional change signal directly triggered by drug perturbation from the pre- and post-perturbation states. In one specific embodiment, a feature interaction network can be integrated into the molecular graph autoencoder. This network can employ a cross-attention mechanism, allowing the first and second transcriptome state features to mutually serve as query conditions, dynamically calculating differential gene expression patterns related to perturbation, while reducing interference from background noise and intrinsic cellular variations. The perturbed transcriptome features obtained through this processing are a vectorized representation, reflecting the essential information of the cellular state transition from pre- to post-perturbation, providing a biological basis for subsequent molecular attribute inference.

[0034] The activity representation extraction module is used to determine the global topological representation and local activity representation based on perturbation transcriptome features. The global topological representation is used to characterize the global topological legitimacy of a molecule, while the local activity representation is used to characterize the local functional activity of a molecule. It should be understood that the process of determining the global topological representation and local activity representation here involves mapping biologically-level perturbation features to two orthogonal chemical property spaces.

[0035] Here, global topology is used to characterize the global topological legitimacy of a molecule, i.e., the chemical rules that the molecule must satisfy at the overall structural level, such as atomic valence compliance, ring system stability, and drug-likeness criteria. To achieve this, the activity representation extraction module aligns perturbed transcriptome features with the latent space of a pre-trained graph autoencoder that captures chemical topological constraints, thereby ensuring the feasibility of the derived molecular topology at the synthetic and structural levels.

[0036] Local activity representation is used to characterize the local functional activity of molecules, focusing on encoding chemical substructure information related to interactions with biological targets, such as pharmacophores and functional fragments. The activity representation extraction module can map perturbed transcriptome features to a representation space that can characterize the local chemical environment, such as a sparse vector space represented by Morgan fingerprints, to achieve alignment of chemical substructure features related to known biological activities, thereby ensuring that the generated molecules possess the expected biological functions.

[0037] Here, the fusion module is used to generate molecular condition vectors based on global topological representations and local activity representations. Fusion methods include, but are not limited to, vector concatenation, weighted combination, or adaptive fusion based on neural networks. This fusion process integrates the dual constraints of molecular structure legitimacy and functional specificity. The resulting molecular condition vector provides comprehensive and structured guidance information for subsequent conditional molecule generation tasks, ensuring that the generated molecules possess both syntheticability and targeted biological activity.

[0038] Step 130: Input the molecular condition vector into the conditional molecular generation model to obtain the molecular structure corresponding to the cell perturbation response output by the conditional molecular generation model.

[0039] Specifically, after obtaining the molecular condition vector Then, the molecular condition vector can be input into the conditional molecular generation model to obtain the molecular structure corresponding to the cell perturbation response output by the conditional molecular generation model.

[0040] Here, the conditional molecule generation model is a generative model. The conditional molecule generation model can be based on the architecture of diffusion model, variational autoencoder, generative adversarial network or flow model. The embodiments of the present invention do not specifically limit this.

[0041] Unlike unconditional generative models, the generation process of conditional molecular generative models is strictly constrained by external conditions. In this invention, these external conditions are molecular condition vectors. For example, in an implementation based on a graph diffusion model, these molecular condition vectors can be injected into the conditional molecular generative model at each step of the denoising process through mechanisms such as adaptive layer normalization. This guides the conditional molecular generative model to gradually denoise from a random noise graph, ultimately generating a molecular graph that satisfies specific chemical constraints.

[0042] The molecular structure corresponding to the cellular perturbation response is the novel molecule that will ultimately be generated and possess the desired biological function. This molecular structure can be represented in various forms, such as a molecular diagram representing the connections between atoms and chemical bonds, or a linear representation such as the string "SMILES". This embodiment of the invention does not specifically limit the representation in this way.

[0043] The method provided in this invention obtains the first transcriptome state features of cells before drug perturbation and the second transcriptome state features of cells after drug perturbation; inputs the first and second transcriptome state features into a dual-view molecular domain model to obtain the molecular condition vector output by the dual-view molecular domain model; inputs the molecular condition vector into a conditional molecular generation model to obtain the molecular structure corresponding to the cell perturbation response output by the conditional molecular generation model. By transforming phenotypic functional differences into constraints for chemical molecule design, the conditional molecular generation model can be guided to generate molecular structures that can induce specific cell perturbation responses. This solves the technical problem of traditional drug design relying excessively on the three-dimensional structural information of the target, and enables accurate de novo molecular design even in scenarios with unknown targets or complex multi-pathway diseases. Meanwhile, this method uses molecular condition vectors characterizing changes in cell phenotypes as generation conditions, overcoming the information asymmetry problem in existing transcriptomics methods that only use phenotypes as prediction targets. It also effectively bridges the modal gap between biological signals and chemical structures through a dual-view alignment strategy, solving the problems of information heterogeneity and semantic gap between biological transcriptomes and chemical structures. This enables a reliable mapping from functional characterization to structural design, significantly improving the feasibility, accuracy, and generalization ability of transcriptome-guided de novo drug molecule design.

[0044] Based on the above embodiments, determining the perturbation transcriptome features characterizing the cellular perturbation response based on the first transcriptome state features and the second transcriptome state features includes: Step 210: Perform feature interaction on the first transcriptome state features and the second transcriptome state features to obtain the first interaction feature and the second interaction feature; Step 220: Fuse the first interaction feature and the second interaction feature to obtain the perturbed transcriptome features.

[0045] Specifically, the steps by which the molecular graph autoencoder determines perturbation transcriptome features characterizing cellular perturbation responses based on the first transcriptome state features and the second transcriptome state features may include: First, the characteristics of the first transcriptome state. Second transcriptome state characteristics Feature interactions are performed to obtain first and second interaction features. These first and second interaction features are then fused to obtain perturbed transcriptome features. .

[0046] In a preferred embodiment, the aforementioned feature interaction process aims to accurately identify and separate the causal signal directly caused by drug perturbation from the complex changes between the two states before and after the perturbation. This process can be implemented by one or more stacked interaction modules, in which a symmetrical cross-attention mechanism can be employed. For example, the first transcriptome state feature can be used as the query, and the second transcriptome state feature as the key and value. Through attention calculation, a feature representation focusing on the post-perturbation state can be obtained, which can be used as part of the first interaction feature. Symmetrically, the second transcriptome state feature can be used as the query, and the first transcriptome state feature as the key and value, to obtain a feature representation focusing on the pre-perturbation state, which is used as part of the second interaction feature. Subsequently, these first and second interaction features can be further refined internally through a self-attention layer to enhance the feature dependencies within their respective states.

[0047] The fusion process aims to integrate the first and second interaction features, after interaction and refinement, into a unified and compact perturbation transcriptome feature. This process can be implemented by an adaptive fusion unit, for example, through a learnable gating mechanism or by summing after a simple linear transformation, to integrate the information from the two interaction features.

[0048] The method provided in this invention, by performing bidirectional feature interaction and fusion on the transcriptome state features before and after perturbation, rather than simple difference or splicing, can force a dual-view molecular domain model to dynamically align the two transcriptional states and highlight drug-driven differential gene expression patterns. This more effectively extracts precise causal perturbation features directly related to therapeutic effects from complex cellular background noise, significantly improving the quality and purity of the conditional signals on which downstream molecular generation tasks depend.

[0049] Based on the above embodiments, the dual-view molecular domain model is obtained by iteratively executing the following steps until a preset iteration termination condition is met: Step 310: Obtain the sample perturbation transcriptome feature set and the initial dual-view molecular domain model; the initial dual-view molecular domain model includes an initial molecular graph autoencoder and an initial activity representation extraction module; the sample perturbation transcriptome feature set includes multiple sample perturbation transcriptome features, the sample perturbation transcriptome features include a first sample perturbation transcriptome feature and a second sample perturbation transcriptome feature; the initial molecular graph autoencoder includes an encoder and a decoder; the initial activity representation extraction module includes an initial local activity representation extraction module; Step 320: Input the sample perturbed transcriptome features into the initial molecular map autoencoder, whereby the encoder encodes the sample perturbed transcriptome features to obtain a latent code, and the decoder decodes the latent code to obtain a molecular map. Step 330: Determine the global topological alignment loss based on the molecular graph and the latent encoding; Step 340: Obtain the target Morgan fingerprint corresponding to the perturbed transcriptome features of the sample; Step 350: Input the sample perturbed transcriptome features into the initial local activity representation extraction module to obtain the predicted continuous fingerprint representation output by the initial local activity representation extraction module; Step 360: Based on the predicted continuous fingerprint representation and the target Morgan fingerprint, determine the local bioactivity alignment loss; Step 370: Determine the target loss based on the global topology alignment loss and the local bioactivity alignment loss; Step 380: Update the model parameters of the initial dual-view molecular domain model based on the target loss.

[0050] Specifically, the two-view molecular domain model can be obtained through iterative training. A complete training process aims to teach the two-view molecular domain model how to accurately map biological perturbation signals into effective constraints in the chemical space. This iterative training process can repeatedly execute the following steps until a preset termination condition is met, such as reaching a preset number of training rounds or the performance of the two-view molecular domain model on the validation set no longer improving: Step 1: Obtain the sample perturbation transcriptome feature set and the initial two-view molecular domain model. The sample perturbation transcriptome feature set is the training dataset, containing multiple sample perturbation transcriptome features. Each sample perturbation transcriptome feature is calculated based on a pair of transcriptome state features before and after perturbation; for example, it may include a first sample perturbation transcriptome feature and a second sample perturbation transcriptome feature. The initial two-view molecular domain model refers to the model whose parameters are randomly initialized or obtained through pre-training before training begins. Its structure includes an initial molecular graph autoencoder and an initial activity representation extraction module. The initial molecular graph autoencoder includes an encoder and a decoder; the initial activity representation extraction module includes an initial local activity representation extraction module.

[0051] Step 2: The perturbed transcriptome features of the sample are input into the initial molecular graph autoencoder. The encoder encodes the perturbed transcriptome features to obtain the latent code, which is then decoded by the decoder to obtain the molecular graph. This process aims to establish a mapping relationship between the perturbed transcriptome features and the global topological structure of the molecule. The latent code is a low-dimensional vector representation obtained by compressing the perturbed transcriptome features of the sample; it captures the core essence of the input information. The molecular graph reflects the graph structure data representation of the molecule, with atoms as nodes and chemical bonds as edges, and is reconstructed by the decoder based on the latent code.

[0052] Step 3: Determine the global topological alignment loss based on the molecular graph and latent encoding. The global topological alignment loss is used to constrain the generated molecular structure to be topologically valid.

[0053] Step 4: Obtain the target Morgan fingerprint corresponding to the perturbed transcriptome features of the sample. The target Morgan fingerprint is the Morgan fingerprint representation of the real molecules associated with the perturbed transcriptome features of the sample, serving as a supervisory signal for local biological activity alignment.

[0054] Step 5: Input the perturbed transcriptome features of the sample into the initial local activity representation extraction module to obtain the predicted continuous fingerprint representation output by the initial local activity representation extraction module.

[0055] Step Six: Based on the predicted continuous fingerprint representation and the target Morgan fingerprint, determine the local bioactivity alignment loss. This local bioactivity alignment loss is used to constrain the generated molecules to have specific local functional activities.

[0056] Step 7: Based on the global topology alignment loss and the local bioactivity alignment loss, determine the target loss. The formula for the target loss is as follows: ; in, Indicates target loss. This represents the global topology alignment loss. This indicates a loss of localized bioactivity alignment. It is a weighted hyperparameter used to balance the importance of two views.

[0057] Step 8: Based on the target loss, update all trainable model parameters of the initial dual-view molecular domain model using the backpropagation algorithm and optimizer.

[0058] The method provided in this invention designs a dual-objective loss function, including global topological alignment and local bioactivity alignment, and jointly optimizes it during iterative training. This enables the dual-view molecular domain model to learn a unified potential representation space that balances the syntheticity of molecular structures with the specificity of biological functions. This training method utilizes robust prior knowledge from the field of chemistry to effectively bridge the cross-modal gap between biological signals and chemical structures, solving the problems of generation failure or functional mismatch that may result from direct mapping, and ensuring the dual high quality of molecules generated by the final dual-view molecular domain model.

[0059] Based on the above embodiments, step 360 includes: Step 3601: Obtain the positive sample mask and the negative sample mask; wherein, the positive sample mask indicates the valid positions in the target Morgan fingerprint, and the negative sample mask indicates the missing positions in the target Morgan fingerprint; Step 3602: Determine the first loss based on the positive sample mask, the predicted continuous fingerprint representation, and the target Morgan fingerprint; Step 3603: Determine the second loss based on the negative sample mask and the predicted continuous fingerprint representation; Step 3604: Determine the sparse regression loss based on the first loss and the second loss; Step 3605: Determine the contrast loss based on the predicted continuous fingerprint representation and the target Morgan fingerprint; Step 3606: Determine the local bioactivity alignment loss based on the sparse regression loss and the contrast loss.

[0060] Specifically, firstly, positive sample masks and negative sample masks can be obtained. Among them, the positive sample mask... It is a binary mask used to indicate valid locations in the target Morgan fingerprint with values ​​greater than 0, i.e., specific chemical substructures or functional groups actually present in the molecule. Negative sample mask. It is a binary mask used to indicate the missing locations of a value of 0 in the target Morgan fingerprint, representing structural features that do not appear in the molecule.

[0061] Secondly, based on the positive sample mask, the predicted continuous fingerprint representation, and the target Morgan fingerprint, the first loss is determined, and the formula for the first loss is as follows: ; in, denoted as the first loss; W is the adaptive weight matrix assigned to the positive sample mask, used to balance the influence of substructures of different frequencies; A represents the predicted continuous fingerprint representation, i.e., the high-dimensional fingerprint representation that is attempted to be reconstructed; B represents the target Morgan fingerprint. This represents the positive sample mask.

[0062] Furthermore, a second loss can be determined based on the negative sample mask and the predicted continuous fingerprint representation. The formula for the second loss is as follows: ; in, Let A represent the second loss, and let A represent the predicted continuous fingerprint representation. This represents the negative sample mask.

[0063] Furthermore, the sparse regression loss can be determined based on the first loss and the second loss. The formula for the sparse regression loss is as follows: ; in, denoted as sparse regression loss, W is an adaptive weight matrix assigned to the positive sample mask to balance the influence of substructures of different frequencies; A represents the predicted continuous fingerprint representation, i.e., the high-dimensional fingerprint representation that is attempted to be reconstructed; B represents the target Morgan fingerprint. Indicates a positive sample mask; α represents the negative sample mask; α is a hyperparameter that controls the penalty strength of the negative sample mask.

[0064] Then, the contrast loss can be determined based on the predicted continuous fingerprint representation and the target Morgan fingerprint. Contrast loss can be based on label mask-based cross-entropy loss, by introducing a sparse regularization penalty for the zero position of the fingerprint, in order to further improve the model's ability to distinguish between positive and negative samples.

[0065] Finally, the local bioactivity alignment loss can be determined based on sparse regression loss and contrast loss. The formula for the local bioactivity alignment loss is as follows: ; in, This indicates a loss of localized bioactivity alignment. Indicates comparative loss, This represents the sparse regression loss.

[0066] The method provided in this invention, by integrating sparse regression loss and contrast loss to calculate specific local bioactivity alignment loss, can effectively solve the training challenges posed by the high-dimensional sparsity of Morgan fingerprints. This method not only ensures that the two-view molecular domain model can accurately reconstruct the key functional groups (positive samples) actually present in the molecule through weighted regression, but also prevents the two-view molecular domain model from falling into the trap of simply setting all predicted outputs to zero by explicitly penalizing zero-value positions. This forces the model to learn the fine-grained local chemical environment crucial to bioactivity, greatly improving the accuracy of the generated molecules in terms of functional specificity.

[0067] Based on the above embodiments, step 330 includes: Step 3301: Based on the molecular graph and the latent encoding, determine the lower bound loss of evidence; Step 3302: Based on the mean of the sample perturbation transcriptome feature set and the mean of the latent code corresponding to each sample perturbation transcriptome feature, as well as the variance of the sample perturbation transcriptome feature set and the variance of the latent code corresponding to each sample perturbation transcriptome feature, determine the feature distribution alignment loss. Step 3303: Determine the global topology alignment loss based on the evidence lower bound loss and the feature distribution alignment loss.

[0068] Specifically, firstly, the lower bound loss of evidence can be determined based on the molecular diagram and the latent coding, as shown in the following formula: ; in, Indicates the loss of the lower limit of evidence. Representing a molecular diagram, Represents latent encoding, Indicates encoder, Indicates decoder, The KL divergence (KL divergence) measures the difference between the latent coding distribution of the encoder output and a pre-defined prior distribution, such as the standard normal distribution. The differences between them It is a weight hyperparameter; This represents the latent encoding in the encoder output. Under the distribution, the log-likelihood of the decoder reconstruction Seeking expectations.

[0069] Then, based on the mean of the sample perturbed transcriptome feature set and the mean of the latent code corresponding to each sample perturbed transcriptome feature, as well as the variance of the sample perturbed transcriptome feature set and the variance of the latent code corresponding to each sample perturbed transcriptome feature, the feature distribution alignment loss can be determined, as shown in the following formula: ; in, This represents the feature distribution alignment loss. This represents the mean of the set of features of the perturbed transcriptome in the sample. This represents the variance of the set of features of the perturbed transcriptome. This represents the mean of the latent coding corresponding to the perturbed transcriptome features of each sample. This represents the variance of the latent coding corresponding to the perturbed transcriptome features of each sample.

[0070] Here, the feature distribution alignment loss is used to reflect the degree of alignment between the distribution of perturbed transcriptome features representing biological signals and the distribution of the underlying encoding representing chemical topological information. The feature distribution alignment loss forces the dual-view molecular domain model to learn the semantic alignment relationship between the two modalities by penalizing the difference in the mean and variance of the two feature distributions.

[0071] Among them, the mean of the sample perturbed transcriptome feature set This information reflects the overall statistical distribution of all perturbed transcriptome features in this batch of training samples. The mean of the latent codes corresponding to the perturbed transcriptome features of each sample. It is used to reflect the overall statistical distribution information of the potential codes obtained after the molecular map corresponding to the batch of samples is encoded.

[0072] Finally, based on the evidence lower bound loss and the feature distribution alignment loss, the global topology alignment loss is determined, as shown in the following formula: ; in, This represents the global topology alignment loss. The standard loss generated for the constraint graph, i.e., the lower bound loss of evidence. This represents the feature distribution alignment loss.

[0073] The method provided in this invention explicitly decomposes the global topological alignment loss into an evidence lower bound loss and a feature distribution alignment loss. This invention not only ensures the rationality of the basic chemical structure of the generated molecule through the evidence lower bound loss, but also forcibly narrows the distance between the biological signal domain and the chemical topological domain at the distribution level through the feature distribution alignment loss. This design makes the cross-modal mapping relationship learned by the model more robust and generalized, effectively solving the mapping instability problem caused by large modal differences, and ensuring that the conversion from biological signals to topologically valid molecular structures is reliable and semantically consistent.

[0074] Based on the above embodiments, step 110 includes: Step 1101: Obtain the first single-cell transcriptome data and the second single-cell transcriptome data; Step 1102: Project the first single-cell transcriptome data and the second single-cell transcriptome data onto a low-dimensional manifold space to obtain the first projected cell embedding and the second projected cell embedding. Step 1103: Based on the cell cycle of the cells, the first projected cell embedding and the second projected cell embedding are grouped respectively to obtain the first structured feature matrix and the second structured feature matrix; Step 1104: Perform local pooling aggregation within each group of the first structured feature matrix and the second structured feature matrix, and perform hierarchical sampling and recombination on the results of the local pooling aggregation to obtain the first transcriptome state features and the second transcriptome state features.

[0075] Specifically, firstly, first and second single-cell transcriptome data can be obtained. The first and second single-cell transcriptome data are sets of original gene expression data measured from a large number of single cells before and after drug perturbation, respectively, and are usually represented as high-dimensional sparse counting matrices.

[0076] Secondly, the first and second single-cell transcriptome data can be projected onto a low-dimensional manifold space to obtain the first and second projected cell embeddings, respectively. This step aims to reduce the dimensionality and noise of the original high-dimensional sparse data.

[0077] Among them, low-dimensional manifold space is a vector space with lower dimensions that can capture core biological information.

[0078] Here, the first projected cell embedding and the second projected cell embedding are the vector representations of the first and second single-cell transcriptome data in the low-dimensional manifold space, respectively. Specifically, large-scale pre-trained base models such as SCimilarity can be used to project high-dimensional, sparse raw expression profiles into a biologically information-dense low-dimensional latent manifold space to alleviate the curse of dimensionality and dropout events.

[0079] Furthermore, the first projected cell embedding and the second projected cell embedding can be grouped based on the cell cycle of the cell to obtain the first structured feature matrix and the second structured feature matrix.

[0080] Here, the cell cycle is used to reflect different stages of cell proliferation, such as G1 phase, S phase, and G2 / M phase. Grouping cells according to their cell cycle stage creates biologically coherent cell clusters. The first and second structured feature matrices are matrix structures formed by embedding cells into groups according to their cell cycle, where each row (or column) of the submatrix corresponds to a specific cell cycle subgroup.

[0081] Finally, local pooling aggregation can be performed within each group of the first and second structured feature matrices, and the results of local pooling aggregation are then subjected to hierarchical sampling and recombination to obtain the first and second transcriptome state features. This step is the Cycle-Stratified Structured Aggregation strategy. It should be noted that the system will hierarchically sample a fixed number (e.g., N=128) of cell feature vectors from each subpopulation based on the proportion of each cell cycle subpopulation in the overall cell population. Subsequently, these sampled feature vectors are aggregated, such as through pooling operations, to generate a fixed-dimensional conditional embedding that characterizes the overall cell population distribution state. This conditional embedding will serve as conditional information reflecting the cell population cycle state and will be input into subsequent modules.

[0082] If the input data for the dual-view molecular domain model is a tissue (Bulk) sample, which does not contain single-cell-level cycle heterogeneity information, then the above steps of grouping, sampling, and generating conditional embedding based on single-cell cycle will be skipped, and the system will directly enter the next stage of the processing flow.

[0083] Local pooling aggregation is performed within these biologically coherent clusters, for example, by averaging. Then, stratified sampling and recombination are performed according to the proportion of each cell cycle subpopulation in the whole, ultimately yielding a fixed-dimensional feature vector that represents the overall cell population distribution characteristics while preserving the heterogeneity information of key subpopulations. This vector serves as the first and second transcriptome state features. In other words, the first and second transcriptome state features are used to simultaneously reflect the overall transcriptome state of the cell population and the heterogeneity information of its key biological subpopulations within a fixed-dimensional representation.

[0084] The method provided in this invention, by introducing a structured denoising and aggregation mechanism, can effectively handle highly challenging single-cell transcriptome data. This mechanism acts as a structured denoiser, solving the fundamental problem of loss of drug response heterogeneity information caused by conventional global averaging methods folding complex cell populations into a single mean vector. While smoothing technical random noise, this method strictly preserves the distribution variance and subpopulation heterogeneity pharmacological signals required for accurate drug design, thus providing downstream models with a much richer and more accurate input signal than traditional methods, significantly improving the ability and robustness of function-driven molecule generation at single-cell resolution.

[0085] Based on the above embodiments, the step of determining the perturbation transcriptome features characterizing the cellular perturbation response based on the first transcriptome state features and the second transcriptome state features further includes: Step 410: Determine the similarity coefficient between the perturbed transcriptome features and the fingerprints of each compound in the preset compound fingerprint library; Step 420: Based on the similarity coefficient, candidate compounds are screened from the compound fingerprint database.

[0086] Specifically, first, the similarity coefficient between the perturbed transcriptome features and the fingerprints of each compound in the pre-defined compound fingerprint database is determined, and then candidate compounds are screened from the compound fingerprint database based on the similarity coefficient.

[0087] In this embodiment, perturbed transcriptome features can be used not only for de novo generation but also as a functional query template for searching within an existing chemical space. The similarity coefficient is a numerical value reflecting the degree of match between the ideal chemical properties encoded by the perturbed transcriptome features and the chemical properties of a real compound in a compound fingerprint database. The similarity coefficient can be obtained by calculating the distance or similarity between the perturbed transcriptome features and each compound fingerprint (such as a Morgan fingerprint) in the compound fingerprint database; for example, the Tanimoto similarity coefficient can be used.

[0088] Based on the calculated similarity coefficients, the system can sort all compounds in the compound fingerprint database and select the compounds with the highest similarity coefficients as candidate compounds. For example, a Top-K nearest neighbor search can be performed. When a threshold is set, such as k=10, the system can directly select a sub-library of candidate clinical compounds with extremely high structural and functional fidelity.

[0089] The method provided in this invention cleverly extends the de novo generative framework into a highly efficient zero-sample virtual screening tool by adding a similarity-based search screening step. Utilizing learned perturbed transcriptome features rich in biological functional information as probes, it can rapidly identify potentially effective candidate drugs in large compound libraries without any additional model training. This method enables zero-sample rapid identification of gene inhibitors and other drugs, especially in time-sensitive therapeutic applications, significantly shortening the drug discovery cycle and providing strong technical support for drug retargeting and rapid response to emerging diseases.

[0090] This invention completely eliminates the absolute dependence on three-dimensional protein target structural information, pioneering a new paradigm of drug design guided by system-level functional outcomes (i.e., cell phenotypic variation). This method can achieve precise de novo molecular design directly based on the overall cellular response map in complex situations where the target structure is unknown, the target is not drugable, or the disease phenotype involves multiple pathway dysregulations. By introducing a dual-view molecular domain alignment strategy, the dual-view molecular domain model effectively solves the pathological inverse mapping problem from macroscopic biological signals to microscopic atomic connections, achieving high-fidelity structure generation even under conditions of significant modal differences. Experimental verification shows that the generated molecules perform excellently in key indicators such as effectiveness, novelty, and Fréchet ChemNet distance, avoiding the simple memorization of the training set scaffold by the dual-view molecular domain model and achieving true chemical space exploration. Furthermore, the system based on the periodic hierarchical heterogeneity aggregation module exhibits strong noise robustness, maintaining robust distribution alignment and feature fidelity even in sparse single-cell data environments where traditional pseudo-tissue averaging methods are prone to failure. Most importantly, despite the complete absence of protein structure information during the training phase, the molecules generated by this invention exhibited excellent binding affinity (up to -9.164 kcal / mol) in zero-sample generation tests targeting core targets such as EGFR, significantly surpassing clinical benchmark inhibitors (such as erlotinib, -7.302 kcal / mol). This demonstrates that the dual-view molecular domain model can extract high-resolution structural constraint information from transcriptome features and successfully generate novel molecular architectures with superior targeting through a "skeleton transition" strategy.

[0091] In industrial deployment and system expansion, this solution can also adopt the following alternative technical paths to achieve functional extension or resource optimization based on actual needs. In the transcriptome data processing stage, if computational resources limit the use of the SCimilarity manifold projection and periodic hierarchical aggregation module based on large-scale pre-training, a de novo trained multilayer perceptron or convolutional neural network can be used to perform nonlinear feature compression. The simplest solution is to directly use the naive averaging method for aggregation, but this requires accepting a trade-off of reduced functional relevance of generated molecules. Regarding the construction of a functional screening closed loop, in addition to full de novo generation, for applications with extremely high timeliness requirements, the extracted perturbed transcriptome features can be used only as search criteria, combined with traditional quantitative evaluation descriptors and fingerprint algorithms, to perform efficient virtual screening of ultra-large-scale existing molecular libraries. This allows the system to be used as a high-performance prediction and screening tool, still demonstrating significant translational medical value.

[0092] Based on any of the above embodiments Figure 2 This is the second schematic diagram of the molecular generation method provided by the present invention, as shown below. Figure 2As shown, the architecture can be logically divided into two parts: a functional space and a molecular space, which together constitute a complete mapping process from biological function to chemical generation. Specifically, the method first takes the acquired first and second transcriptome state features as inputs in the functional space and feeds them into a molecular autoencoder. This molecular autoencoder extracts signals characterizing the net effect of drug perturbation through feature interactions, such as cross-attention mechanisms, and passes them to the activity representation extraction module. The activity representation extraction module uses a dual-view alignment strategy to map the biological signals in parallel into two complementary chemical representations: one is a global topological representation for characterizing the legitimacy of molecular structure, and the other is a local activity representation for characterizing the specificity of molecular function. These two representations together constitute a comprehensive perturbation transcriptome feature z. Subsequently, the perturbation transcriptome feature z is processed into the final molecular conditional vector C by the encoding module and fed into the conditional molecular generation model in the molecular space to guide the de novo generation of molecules.

[0093] In this preferred embodiment, the conditional molecule generation model is a perturbation feature-guided discrete molecular graph diffusion model, whose generation backbone network adopts a graph diffusion Transformer architecture. Its generation process is a reverse denoising process, the opposite of a Markov chain-driven forward process that progressively adds noise to the discrete features (atoms and bond types) of the molecular graph. Specifically, the conditional molecule generation model starts with a purely random noise graph GT, and through a series of iterative denoising steps, finally outputs a deterministic, structurally valid molecular graph G0 at step t=0. In each iteration of the denoising process, the conditional molecule generation model utilizes classifier-free guidance (CFG) technology to deeply inject the perturbation features implied by the molecular conditional vector C into the generation process. Inside the core unit DiT (Diffusion Transformer) of the diffusion model of the conditional molecule generation model, it includes a multi-head self-attention mechanism and a feedforward neural network. The molecular conditional vector C undergoes scale and translation modulation of the hidden features through an adaptive layer normalization (AdaLN) mechanism. This guidance process can be expressed by the following formula: ; Control, in which the guiding intensity scale 's' is used to adjust the importance of the condition. This represents a conditional molecular generation model with the following parameters: , This represents the final generated distribution without a classifier. This represents the noise molecule graph at step t (the current step). This represents the molecular diagram at step t-1. Represents the molecular conditional vector. This indicates a conditionally generated distribution (guided by perturbation features); Represents an unconditionally generated distribution (dependent only on the noise map). ).

[0094] Through this design, compared to unconstrained or generative models based solely on SMILES sequences, the discrete graph diffusion-based method of this invention can directly reconstruct the graph structure iteratively in the three-dimensional topological space, ensuring extremely high structural effectiveness and generative diversity, and perfectly inheriting the high-dimensional multimodal conditional signals extracted from the functional space upstream.

[0095] The molecular generation apparatus provided by the present invention is described below. The molecular generation apparatus described below and the molecular generation method described above can be referred to in correspondence.

[0096] Based on any of the above embodiments, the present invention provides a molecular generation apparatus. Figure 3 This is a schematic diagram of the molecular generation device provided by the present invention, as shown below. Figure 3 As shown, the device includes: The acquisition module 310 is used to acquire the first transcriptome state characteristics of the cells before drug perturbation and the second transcriptome state characteristics of the cells after drug perturbation. The input module 320 is used to input the first transcriptome state features and the second transcriptome state features into the dual-view molecular domain model to obtain the molecular condition vector output by the dual-view molecular domain model. The generation module 330 is used to input the molecular condition vector into the conditional molecule generation model to obtain the molecular structure corresponding to the cell perturbation response output by the conditional molecule generation model. The dual-view molecular domain model includes a molecular graph autoencoder, an activity representation extraction module, and a fusion module. The molecular graph autoencoder is used to determine perturbation transcriptome features characterizing cellular perturbation responses based on the first transcriptome state features and the second transcriptome state features. The activity representation extraction module is used to determine global topological representation and local activity representation based on the perturbation transcriptome features. The global topological representation is used to characterize the global topological legitimacy of the molecule, and the local activity representation is used to characterize the local functional activity of the molecule. The fusion module is used to fuse the global topological representation and the local activity representation to obtain the molecular conditional vector.

[0097] The apparatus provided in this invention acquires the first transcriptome state features of cells before drug perturbation and the second transcriptome state features of cells after drug perturbation; inputs the first and second transcriptome state features into a dual-view molecular domain model to obtain the molecular condition vector output by the dual-view molecular domain model; inputs the molecular condition vector into a conditional molecular generation model to obtain the molecular structure corresponding to the cell perturbation response output by the conditional molecular generation model. By transforming phenotypic functional differences into constraints for chemical molecule design, the conditional molecular generation model can be guided to generate molecular structures that can induce specific cell perturbation responses. This solves the technical problem of traditional drug design relying excessively on the three-dimensional structural information of the target, and enables accurate de novo molecular design even in scenarios with unknown targets or complex multi-pathway diseases. Meanwhile, this method uses molecular condition vectors characterizing changes in cell phenotypes as generation conditions, overcoming the information asymmetry problem in existing transcriptomics methods that only use phenotypes as prediction targets. It also effectively bridges the modal gap between biological signals and chemical structures through a dual-view alignment strategy, solving the problems of information heterogeneity and semantic gap between biological transcriptomes and chemical structures. This enables a reliable mapping from functional characterization to structural design, significantly improving the feasibility, accuracy, and generalization ability of transcriptome-guided de novo drug molecule design.

[0098] Based on any of the above embodiments, determining the perturbation transcriptome features characterizing the cellular perturbation response based on the first transcriptome state features and the second transcriptome state features includes: The first transcriptome state features and the second transcriptome state features are subjected to feature interaction to obtain the first interaction feature and the second interaction feature; The perturbation transcriptome features are obtained by fusing the first interaction feature and the second interaction feature.

[0099] Based on any of the above embodiments, a training module is further included, wherein the training module specifically includes: A model acquisition module is used to acquire a set of sample perturbation transcriptome features and an initial dual-view molecular domain model. The initial dual-view molecular domain model includes an initial molecular graph autoencoder and an initial activity representation extraction module. The set of sample perturbation transcriptome features includes multiple sample perturbation transcriptome features, including a first sample perturbation transcriptome feature and a second sample perturbation transcriptome feature. The initial molecular graph autoencoder includes an encoder and a decoder. The initial activity representation extraction module includes an initial local activity representation extraction module. The molecular graph module is used to input the sample perturbed transcriptome features into the initial molecular graph autoencoder, whereby the encoder encodes the sample perturbed transcriptome features to obtain a latent code, and the decoder decodes the latent code to obtain the molecular graph. A global loss determination module is used to determine the global topological alignment loss based on the molecular graph and the latent encoding; The target Morgan fingerprint acquisition module is used to acquire the target Morgan fingerprint corresponding to the perturbed transcriptome features of the sample; The prediction module is used to input the sample perturbed transcriptome features into the initial local activity representation extraction module to obtain the predicted continuous fingerprint representation output by the initial local activity representation extraction module; A local loss determination module is used to determine the local bioactivity alignment loss based on the predicted continuous fingerprint representation and the target Morgan fingerprint; The target loss determination module is used to determine the target loss based on the global topology alignment loss and the local bioactivity alignment loss; The parameter update module is used to update the model parameters of the initial dual-view molecular domain model based on the target loss.

[0100] Based on any of the above embodiments, the local loss determination module is specifically used for: Obtain a positive sample mask and a negative sample mask; wherein, the positive sample mask indicates the valid positions in the target Morgan fingerprint, and the negative sample mask indicates the missing positions in the target Morgan fingerprint; Based on the positive sample mask, the predicted continuous fingerprint representation, and the target Morgan fingerprint, a first loss is determined; The second loss is determined based on the negative sample mask and the predicted continuous fingerprint representation; Based on the first loss and the second loss, determine the sparse regression loss; Based on the predicted continuous fingerprint representation and the target Morgan fingerprint, determine the contrast loss; The local bioactivity alignment loss is determined based on the sparse regression loss and the contrast loss.

[0101] Based on any of the above embodiments, the global loss determination module is specifically used for: Based on the molecular diagram and the latent encoding, the lower bound loss of evidence is determined; Based on the mean of the sample perturbation transcriptome feature set and the mean of the latent code corresponding to each sample perturbation transcriptome feature, as well as the variance of the sample perturbation transcriptome feature set and the variance of the latent code corresponding to each sample perturbation transcriptome feature, the feature distribution alignment loss is determined. The global topology alignment loss is determined based on the evidence lower bound loss and the feature distribution alignment loss.

[0102] Based on any of the above embodiments, the acquisition module 310 is specifically used for: Acquire first and second single-cell transcriptome data; The first single-cell transcriptome data and the second single-cell transcriptome data are projected onto a low-dimensional manifold space to obtain the first projected cell embedding and the second projected cell embedding. Based on the cell cycle of the cells, the first projected cell embedding and the second projected cell embedding are grouped to obtain the first structured feature matrix and the second structured feature matrix; Local pooling aggregation is performed within each group of the first structured feature matrix and the second structured feature matrix, and the results of local pooling aggregation are hierarchically sampled and recombined to obtain the first transcriptome state features and the second transcriptome state features.

[0103] Based on any of the above embodiments, a filtering module is further included, wherein the filtering module is specifically used for: Determine the similarity coefficient between the perturbed transcriptome features and the fingerprints of each compound in a pre-defined compound fingerprint database; Candidate compounds are selected from the compound fingerprint database based on the similarity coefficient.

[0104] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4As shown, the electronic device may include: a processor 410, a communications interface 420, a memory 430, and a communications bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other through the communications bus 440. The processor 410 can call logical instructions in the memory 430 to execute a molecular generation method, which includes: acquiring first transcriptome state features of cells before drug perturbation and second transcriptome state features of cells after drug perturbation; inputting the first transcriptome state features and the second transcriptome state features into a dual-view molecular domain model to obtain a molecular condition vector output by the dual-view molecular domain model; inputting the molecular condition vector into a conditional molecular generation model to obtain the molecular structure corresponding to the cell perturbation response output by the conditional molecular generation model; the dual-view molecular domain model includes a molecular graph autoencoder, an activity representation extraction module, and a fusion module; the molecular graph autoencoder is used to determine perturbed transcriptome features characterizing the cell perturbation response based on the first transcriptome state features and the second transcriptome state features; the activity representation extraction module is used to determine a global topological representation and a local activity representation based on the perturbed transcriptome features; the global topological representation is used to characterize the global topological legitimacy of the molecule, and the local activity representation is used to characterize the local functional activity of the molecule; the fusion module is used to fuse the global topological representation and the local activity representation to obtain the molecular condition vector.

[0105] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the molecular generation method provided by the above methods. The method includes: acquiring a first transcriptome state feature of cells before drug perturbation and a second transcriptome state feature of cells after drug perturbation; inputting the first transcriptome state feature and the second transcriptome state feature into a dual-view molecular domain model to obtain a molecular condition vector output by the dual-view molecular domain model; and inputting the molecular condition vector into a conditional molecular generation model to obtain the conditional molecular generation model. The molecular structure corresponding to the cellular perturbation response output by the subgenerator model; the dual-view molecular domain model includes a molecular graph autoencoder, an activity representation extraction module, and a fusion module; the molecular graph autoencoder is used to determine perturbation transcriptome features characterizing the cellular perturbation response based on the first transcriptome state features and the second transcriptome state features; the activity representation extraction module is used to determine global topological representation and local activity representation based on the perturbation transcriptome features; the global topological representation is used to characterize the global topological legitimacy of the molecule, and the local activity representation is used to characterize the local functional activity of the molecule; the fusion module is used to fuse the global topological representation and the local activity representation to obtain the molecular conditional vector.

[0107] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the molecular generation method provided by the methods described above. The method includes: acquiring first transcriptome state features of cells before drug perturbation and second transcriptome state features of the cells after drug perturbation; inputting the first transcriptome state features and the second transcriptome state features into a dual-view molecular domain model to obtain a molecular condition vector output by the dual-view molecular domain model; and inputting the molecular condition vector into a conditional molecular generation model to obtain a cell perturbation response output by the conditional molecular generation model. The corresponding molecular structure; the dual-view molecular domain model includes a molecular graph autoencoder, an activity representation extraction module, and a fusion module; the molecular graph autoencoder is used to determine perturbation transcriptome features characterizing cellular perturbation responses based on the first transcriptome state features and the second transcriptome state features; the activity representation extraction module is used to determine global topological representation and local activity representation based on the perturbation transcriptome features; the global topological representation is used to characterize the global topological legitimacy of the molecule, and the local activity representation is used to characterize the local functional activity of the molecule; the fusion module is used to fuse the global topological representation and the local activity representation to obtain the molecular conditional vector.

[0108] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0109] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating molecules, characterized in that, include: Obtain the first transcriptome state characteristics of the cells before drug perturbation and the second transcriptome state characteristics of the cells after drug perturbation; The first transcriptome state features and the second transcriptome state features are input into the dual-view molecular domain model to obtain the molecular condition vector output by the dual-view molecular domain model. The molecular condition vector is input into the conditional molecular generation model to obtain the molecular structure corresponding to the cell perturbation response output by the conditional molecular generation model. The dual-view molecular domain model includes a molecular graph autoencoder, an activity representation extraction module, and a fusion module; the molecular graph autoencoder is used to determine perturbation transcriptome features characterizing cellular perturbation responses based on the first transcriptome state features and the second transcriptome state features. The activity representation extraction module is used to determine the global topological representation and the local activity representation based on the perturbed transcriptome features; the global topological representation is used to characterize the global topological legitimacy of the molecule, and the local activity representation is used to characterize the local functional activity of the molecule; the fusion module is used to fuse the global topological representation and the local activity representation to obtain the molecular condition vector.

2. The molecular generation method according to claim 1, characterized in that, The determination of perturbation transcriptome features characterizing cellular perturbation responses based on the first transcriptome state features and the second transcriptome state features includes: The first transcriptome state features and the second transcriptome state features are subjected to feature interaction to obtain the first interaction feature and the second interaction feature; The perturbation transcriptome features are obtained by fusing the first interaction feature and the second interaction feature.

3. The molecular generation method according to claim 1, characterized in that, The dual-view molecular domain model is obtained by iteratively executing the following steps until a preset iteration termination condition is met: A sample perturbation transcriptome feature set and an initial dual-view molecular domain model are obtained; the initial dual-view molecular domain model includes an initial molecular graph autoencoder and an initial activity representation extraction module; the sample perturbation transcriptome feature set includes multiple sample perturbation transcriptome features, the sample perturbation transcriptome features include a first sample perturbation transcriptome feature and a second sample perturbation transcriptome feature; the initial molecular graph autoencoder includes an encoder and a decoder; the initial activity representation extraction module includes an initial local activity representation extraction module; The sample perturbed transcriptome features are input into the initial molecular map autoencoder, the encoder encodes the sample perturbed transcriptome features to obtain the latent code, and the decoder decodes the latent code to obtain the molecular map; Based on the molecular graph and the latent encoding, the global topological alignment loss is determined; Obtain the target Morgan fingerprint corresponding to the perturbed transcriptome features of the sample; The sample perturbation transcriptome features are input into the initial local activity representation extraction module to obtain the predicted continuous fingerprint representation output by the initial local activity representation extraction module; Based on the predicted continuous fingerprint representation and the target Morgan fingerprint, determine the local bioactivity alignment loss; The target loss is determined based on the global topology alignment loss and the local bioactivity alignment loss; Based on the target loss, update the model parameters of the initial dual-view molecular domain model.

4. The molecular generation method according to claim 3, characterized in that, The step of determining the local bioactivity alignment loss based on the predicted continuous fingerprint representation and the target Morgan fingerprint includes: Obtain a positive sample mask and a negative sample mask; wherein, the positive sample mask indicates the valid positions in the target Morgan fingerprint, and the negative sample mask indicates the missing positions in the target Morgan fingerprint; Based on the positive sample mask, the predicted continuous fingerprint representation, and the target Morgan fingerprint, a first loss is determined; The second loss is determined based on the negative sample mask and the predicted continuous fingerprint representation; Based on the first loss and the second loss, determine the sparse regression loss; Based on the predicted continuous fingerprint representation and the target Morgan fingerprint, determine the contrast loss; The local bioactivity alignment loss is determined based on the sparse regression loss and the contrast loss.

5. The molecular generation method according to claim 3, characterized in that, The determination of the global topological alignment loss based on the molecular graph and the latent encoding includes: Based on the molecular diagram and the latent encoding, the lower bound loss of evidence is determined; Based on the mean of the sample perturbation transcriptome feature set and the mean of the latent code corresponding to each sample perturbation transcriptome feature, as well as the variance of the sample perturbation transcriptome feature set and the variance of the latent code corresponding to each sample perturbation transcriptome feature, the feature distribution alignment loss is determined. The global topology alignment loss is determined based on the evidence lower bound loss and the feature distribution alignment loss.

6. The molecular generation method according to any one of claims 1 to 5, characterized in that, The acquisition of the first transcriptome state characteristics of cells before drug perturbation and the second transcriptome state characteristics of cells after drug perturbation includes: Acquire first and second single-cell transcriptome data; The first single-cell transcriptome data and the second single-cell transcriptome data are projected onto a low-dimensional manifold space to obtain the first projected cell embedding and the second projected cell embedding. Based on the cell cycle of the cells, the first projected cell embedding and the second projected cell embedding are grouped to obtain the first structured feature matrix and the second structured feature matrix; Local pooling aggregation is performed within each group of the first structured feature matrix and the second structured feature matrix, and the results of local pooling aggregation are hierarchically sampled and recombined to obtain the first transcriptome state features and the second transcriptome state features.

7. The molecular generation method according to any one of claims 1 to 5, characterized in that, The step of determining perturbation transcriptome features characterizing cellular perturbation responses based on the first and second transcriptome state features further includes: Determine the similarity coefficient between the perturbed transcriptome features and the fingerprints of each compound in a pre-defined compound fingerprint database; Candidate compounds are selected from the compound fingerprint database based on the similarity coefficient.

8. A molecular generating apparatus, characterized in that, include: The acquisition module is used to acquire the first transcriptome state characteristics of the cells before drug perturbation and the second transcriptome state characteristics of the cells after drug perturbation. The input module is used to input the first transcriptome state features and the second transcriptome state features into the dual-view molecular domain model to obtain the molecular condition vector output by the dual-view molecular domain model. The generation module is used to input the molecular condition vector into the conditional molecule generation model to obtain the molecular structure corresponding to the cell perturbation response output by the conditional molecule generation model. The dual-view molecular domain model includes a molecular graph autoencoder, an activity representation extraction module, and a fusion module; the molecular graph autoencoder is used to determine perturbation transcriptome features characterizing cellular perturbation responses based on the first transcriptome state features and the second transcriptome state features. The activity representation extraction module is used to determine the global topological representation and the local activity representation based on the perturbed transcriptome features; the global topological representation is used to characterize the global topological legitimacy of the molecule, and the local activity representation is used to characterize the local functional activity of the molecule; the fusion module is used to fuse the global topological representation and the local activity representation to obtain the molecular condition vector.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the molecular generation method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the molecular generation method as described in any one of claims 1 to 7.

Images